Here is a contradiction worth sitting with. Every serious forecaster in 2024 told you the same story: training costs were compounding by an order of magnitude a year, the field could not sustain a dozen frontier labs, and 2026 or 2027 would be the years of consolidation. Instead, the opposite happened. More companies are training strong models than ever, and an increasing number are giving the weights away.
The reason the prediction inverted is buried in a single number that made Josh Wolfe interrupt a speaker and ask him to say it again.
Together AI went from serving 30 billion tokens a month to 400 trillion. That is a 10,000x jump, and it is not an infrastructure curiosity. It is the sound of enterprise procurement quietly rewriting its own decision criteria.
Together AI went from serving 30 billion tokens a month to 400 trillion. That is the sound of enterprise procurement quietly rewriting its own decision criteria.
Last week we established that open weights now carry a majority of production traffic and that the seven highest-volume models on OpenRouter all ship open. This week the story hardens: we have the volume data, and we have a fourth parity model to add to the pile.
Start with what the 400T number actually represents. It is not researchers kicking tires. It is AI-native companies and enterprises moving scaled workloads onto open models — the kind of sustained, high-volume traffic that only shows up when a procurement decision has already been made.
And procurement, it turns out, has switched its buying metric. Not quietly announced it. Quietly switched it.
The old axis was peak benchmark score: who sits at the top of the leaderboard. The new axis is cost-per-successful-task. On that axis, the frontier premium collapses for most of what companies actually run.
Together's own DeepSWE analysis makes the mechanism legible. Kimi K3 Max came close to Fable 5 xhigh on Pass@1 while costing about one-third as much per rollout. The result: 2.8x more solved tasks per dollar.
Read the framing carefully, because it is the whole thesis in one line. "For teams running models at scale, the cost per successful task is often the more useful comparison." Not the peak score. The tasks-per-dollar.
Once you adopt that metric, the leaderboard stops being the map. A model that scores three points lower but costs a third as much doesn't lose — it wins by 2.8x on the number that hits the budget.
GLM-5.2 shows the same physics on coding. According to byteiota, Z.ai's model scored 62.1 on SWE-bench Pro against GPT-5.5's 58.6 — it beats the frontier model outright — while pricing at $1.40 input and $4.40 output versus GPT-5.5's $5 and $30.
That is reportedly roughly six times cheaper, on the industry's most-watched coding benchmark, with the open model ahead on the score too. When both the quality line and the cost line point the same direction, the procurement case stops being a debate.
This is why the consolidation forecast failed. The labs everyone expected to merge realized that building "token machines" is a viable path to value on its own — you do not need the single smartest model if you can serve the commodity middle profitably at scale.
Thinking Machines is the sharpest example. When Mira Murati's company launched in February 2025, almost nobody filed it under "open models company." Now, per Interconnects, its fine-tuning service reportedly makes hundreds of millions a year, and it reportedly ships the strongest open-weight models built in the U.S. — ahead of NVIDIA's Nemotron and Arcee's Trilogy.
The frontier gap, meanwhile, has narrowed to something close to a single release cycle. Per Epoch, its Capabilities Index puts the best open model, Kimi K3, six points behind the closed frontier — with overlapping confidence intervals. Six points is roughly one generation.
Two releases show how fast the interior of the field is now iterating. DeepSeek dropped a V4-Flash update one day after OpenAI cut its smallest model's price by 80%, and, per Simon Willison, the 304-billion-parameter Flash beat models larger than itself — ranking ahead of the 428B MiniMax M3 at $0.14 input and $0.27 output. It outran the field while its own flagship Pro sat unreleased.
Then there is the fourth parity model, and the one that makes the release-cycle claim concrete: Inkling-Small from Thinking Machines.
The name misleads. Per Interconnects, it is a 276-billion-parameter MoE with 12 billion active — "small" only next to the 975B Inkling flagship it was distilled from.
The interesting part is that the student caught the teacher. Trained later, it inherited the flagship's data recipe, then ran on-policy distillation and two more weeks of agentic coding RL — and, per the model card, came out ahead of the bigger model on several reasoning and coding metrics: SWE-bench Verified at 80.2 versus 77.6, Terminal Bench and ARC-AGI-2 both higher.
On Artificial Analysis's more conservative reading, Inkling-Small lands at 40 on the Intelligence Index against the flagship's 41 — a single point, at a fraction of the deployment cost. When a cheaper follow-up nearly matches the flagship it descended from within weeks, "one release cycle" stops being a metaphor and becomes a schedule.
That is four models — Kimi K3 Max, GLM-5.2, DeepSeek V4-Flash, Inkling-Small — telling the same story from different angles. The gap between frontier and open is now measured in a release, not a chasm.
Now the qualification that keeps this honest, because a migration is not a rout.
The frontier still earns its premium — but in a narrow band. Claude Opus 4.8 still leads on the hardest tasks: SWE-Marathon, the ultra-long-horizon benchmark, shows Opus at 26.0 against GLM-5.2's 13.0. If your workflow is multi-day autonomous coding, the frontier holds the edge.
The State of Open Source report draws the same jagged line. Open leads or ties on frontend coding and much agentic terminal work; it contests the middle; it concedes the closed edge on deep reasoning, long-context reliability, and professional-knowledge polish. Per the report, Fable 5 still beats Kimi K3 by 92 Elo on expert-graded professional work.
Per byteiota, GLM-5.2 carries its own asterisks: no vision at launch, and it burns roughly 43,000 output tokens per long-horizon task, nearly double a competitor's — which quietly inflates the real cost-per-task the price sheet advertises.
So the frontier is not obsolete. It is being priced correctly for the first time — reserved for the narrow band where long-horizon autonomy and professional polish genuinely pay for themselves.
But here is the thing about the commodity middle: it is where the tokens are. It is where the money is. Four hundred trillion tokens a month do not route to the hardest 5% of problems — they route to the vast, ordinary interior of production work, and that interior has gone open.
That is why the labs are proliferating instead of merging. There is a viable business in serving the middle, and the middle is enormous.
One thread worth flagging for later. The State of Open Source report argues that as the model layer commoditizes, value migrates up to the agentic harness above it — Semgrep's security team reached the same conclusion, that harness architecture mattered more than model choice. Where that value lands is a market of its own, and it runs elsewhere in this issue.
For now, the takeaway is narrower and firmer. The consolidation everyone booked for 2026 didn't arrive because the industry changed the question. It stopped asking who tops the leaderboard and started asking who solves the task per dollar — and on that question, the frontier's monopoly on the answer is gone.